This session aims to bring together the latest advances in the hybrid approaches to (i) improve our understanding of climate system and its variability, (ii) enhance climate predictions on multiple timescales, and (iii) translate these advances into more reliable climate services for diverse users (e.g., health, energy, agriculture, water).
With these objectives in mind, we welcome contributions on, but not limited to: subsampling and filtering strategies to enhance predictions of climate variability and extremes on different timescales (including process-constrained projections); advanced machine learning (ML) and causal discovery techniques for validation, bias-correction and downscaling of dynamical model outputs; hybrid multimodel ensemble approaches such as supermodelling to improve climate model simulations; transfer learning to leverage climate model outputs and expand ML training datasets; physics-informed ML parametrization of sub-grid processes; hybrid surrogate models that emulate or correct specific components of dynamical models; and impact/service oriented studies that deploy hybrid pipelines to support decision-making, such as hybrid seasonal forecasts and early warning systems based on ML or causal discovery techniques.
Orals: Thu, 7 May, 10:45–12:30 | Room 0.14
Low spatiotemporal resolution of global climate models (GCMs) outputs such as CMIP6 models limits accurate detection of tropical cyclone (TC). Traditional statistical downscaling has difficulties in resolving non-linear relationships among different variables, while dynamical downscaling with regional high-resolution models is computational expensive and often distorts the results due to different dynamics framework with the GCMs. Here, we proposed a deep-learning based Multi-Variable Spatiotemporal Downscaling Generative Adversarial Network (MV-STD-GAN). It simultaneously spatiotemporally downscales five essential variables (sea level pressure, 300hPa/500hPa geopotential height, 10m zonal/meridional wind) closely related to TC detection. Trained on high- and low-resolution ERA datasets, it substantially improves the detection of observed TC and significantly outperforms traditional and other deep learning baselines, when subject to the same detection algorithm. It can also be successfully applied to low-resolution CMIP6 models, detecting TC activities very similar to the corresponding high-resolution models.
How to cite: Yuan, C. and Ye, Y.: A Deep Learning–Based Multi-Variable Spatiotemporal Downscaling Approach for High-Resolution Tropical Cyclone Detection, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-2397, https://doi.org/10.5194/egusphere-egu26-2397, 2026.
Machine learning (ML)-based models have recently demonstrated high skill and computational efficiency, often outperforming conventional physics-based models in weather forecasting and subseasonal prediction. While prior efforts have assessed their ability to capture atmospheric dynamics at the synoptic scale, their performance across broader timescales and under out-of-distribution forcing remains insufficiently understood but essential criterion for establishing their credibility in Earth system science.
In this study, we design three idealized test cases to evaluate the Neural General Circulation Model (NeuralGCM), a hybrid model that couples a dynamical core with ML-based physical parameterizations. The test casts span synoptic-scale phenomena, interannual variability, and out-of-distribution forcings via uniform warmings. We benchmark NeuralGCM against observations and conventional physics-based Earth system models (ESMs). At the synoptic scale, NeuralGCM captures the evolution and propagation of extratropical cyclones with performance comparable to ESMs. At the interannual scale, when forced by El Niño-Southern Oscillation sea surface temperature (SST) anomalies, NeuralGCM successfully reproduces associated teleconnection patterns but exhibits deficiencies in capturing nonlinear response. Under out-of-distribution uniform-warming forcings, NeuralGCM simulates similar responses in global-average temperature and precipitation and reproduces large-scale tropospheric circulation features similar to those in ESMs. Notable weaknesses include overestimating the tracks and spatial extent of extratropical cyclones, and biases in the teleconnected wave train triggered by tropical SST anomalies. Furthermore, its simulated temperature responses near the tropopause and in the stratosphere under uniform warming simulations deviate from those in physics-based models, likely due to the biases in vertical temperature advection by the residual circulation.
Despite these limitations, NeuralGCM exhibits credible responses across all test cases and performs comparably to both observations and physics-based ESMs. These results suggest that hybrid models like NeuralGCM, which integrate dynamical cores with ML physics, offer a promising path toward the next generation of ML-based ESMs.
How to cite: Chen, Z., Leung, L. R., Zhou, W., Lu, J., Lubis, S. W., Liu, Y., Chang, J., Harrop, B. E., Wang, Y., Yang, M., Zhang, G., and Qian, Y.: Hierarchical Testing of a Hybrid Machine Learning-Physics Global Atmosphere Model, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-22911, https://doi.org/10.5194/egusphere-egu26-22911, 2026.
Arctic sea ice has declined markedly over the past four decades, bringing new opportunities for commercial shipping and eco-tourism while elevating risks to shipping safety, which highlights the urgent demand for a seamless and skillful sea-ice prediction system. Current sea-ice prediction approaches fall into two main categories: statistical and dynamical models. The former (including machine learning) show promising performance in sea-ice concentration prediction but lack sufficient physical constraints due to their pure data-driven nature; the latter (including Earth system models) are physically grounded with strong consistency and multivariable coherence, yet suffer from high computational costs, imperfect parameterizations and accumulated forecast errors. To overcome these limitations, we propose a hybrid framework that integrates the complementary strengths of machine learning and dynamical models. Specifically, we develop an interactive ensemble prediction system, termed SUPER, which couples the machine learning model Ice-kNN with the Norwegian Climate Prediction Model (NorCPM) to enable recursive information exchange between the two models. Within SUPER, Ice-kNN and NorCPM run in parallel and exchange information at regular intervals (e.g., weekly), allowing mutual adjustment of their predictions. We conduct seasonal hindcasts for the period 2000–2023 and evaluate the daily sea-ice prediction skill of SUPER against that of the standalone models. Preliminary results indicate that the hybrid system substantially reduces errors in sea-ice concentration and sea-ice edge predictions for NorCPM, while yielding improvements for Ice-kNN. Further tuning and evaluation are ongoing, and updated results will be presented.
How to cite: Lin, Y., Wang, Y., Min, C., and Yang, Q.: An Interactive Hybrid Framework Coupling Machine Learning and Dynamical Models for Arctic Sea Ice Prediction, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-22955, https://doi.org/10.5194/egusphere-egu26-22955, 2026.
Mesoscale Convective Systems (MCSs) account for most of the rainfall over the Sahel and contribute substantially to extreme precipitation events (EPEs) that cause flooding in West Africa. However, the low density of ground-based observation networks limits the accurate monitoring and characterization of these events. Regional climate models, such as the Weather Research and Forecasting (WRF) model, help to compensate for this observational gap but still exhibit significant biases in simulating extreme precipitation.
This study aims to reduce these biases by integrating the WRF model into a Supermodelling framework based on the dynamic combination of multiple model configurations, designed to exploit their complementary strengths.
Two supermodels, SUPPERT-WRFA and SUPPERT-WRFB, are developed from three distinct WRF configurations that mainly differ in their convection schemes. SUPPERT-WRFA is trained using satellite-based IMERG precipitation, while SUPPERT-WRFB relies on atmospheric dynamical variables (U, V, and T) from ERA5. In both cases, the training strategy is based on a metric combining spatial correlation and root-mean-square error.
The performance of the supermodels is evaluated for the simulation of EPEs in West Africa through a case study of an extreme precipitation event that occurred on 5 September 2020 in Senegal. The results show a significant improvement in the representation of extreme precipitation compared to individual WRF configurations, highlighting the strong potential of Supermodelling to improve extreme event prediction in regions with sparse observational coverage.
How to cite: Dieng, A. L., Keenlyside, N., Koseki, S., and Schevenhoven, F.: WRF Supermodelling for Improved Simulation of Extreme Precipitation in West Africa, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-22842, https://doi.org/10.5194/egusphere-egu26-22842, 2026.
The North Atlantic exhibits prominent multidecadal variability affecting climate impacts across Europe and the Arctic. Yet, separating internally generated variability from externally forced components remains challenging, especially when univariate indices are used and when models exhibit a low signal-to-noise ratio and biases in coupled feedbacks. Here we present a hybrid statistical–dynamical framework to identify and interpret a predictable coupled multidecadal mode linking the North Atlantic Oscillation (NAO), the Atlantic Meridional Overturning Circulation (AMOC), and Atlantic Multidecadal Variability (AMV), with implications for Arctic sea-ice variability.
We combine (i) long climate-model experiments (preindustrial control and transient forced integrations, including volcanic-only forcing) with (ii) subsampling/filtering using joint multivariate Singular Spectrum Analysis (MSSA) applied to a physically motivated set of fields spanning the stratosphere–troposphere–ocean system. Treating NAO/AMV as an inherently coupled multivariate process enables a clearer separation of signal from noise and isolates an oscillatory mode with a characteristic multidecadal timescale that emerges in unforced control conditions. External forcing primarily modulates the mode’s amplitude and apparent period rather than generating it: volcanic perturbations project efficiently onto ocean circulation and can intermittently excite the coupled state.
To interpret these results, we use a minimal conceptual model of a damped coupled NAO–AMV oscillator under periodic and stochastic forcing. The model demonstrates that multidecadal oscillations in unforced control simulations can be sustained by white-noise atmospheric variability, and clarifies how periodic forcing shape phase, period, and amplitude modulation.
How to cite: Omrani, N.-E. and Keenlyside, N.: A hybrid statistical–dynamical framework for a coupled multidecadal NAO–AMV oscillation, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-22174, https://doi.org/10.5194/egusphere-egu26-22174, 2026.
While future emission pathways are the primary source of uncertainty in long-term global climate projections, internal climate variability dominates near-term uncertainty at the regional scale. Reducing this source of uncertainty is crucial, as it aligns with the planning horizons of stakeholders in climate-sensitive sectors. To address this challenge, decadal forecasts aim to reduce this uncertainty by initializing the climate model simulations from estimates of the observed state of the climate system, in order to phase the temporal evolution of the simulated and observed modes of climate variability. However, decadal forecasts are also subject to drift due to the initialisation shock arising from mismatch between biased models and assimilated observational estimates. In this study, we propose a novel decadal forecasting method based on a process-based constraint approach. This approach aims to align model internal variability with observations by selecting the member closest to the observed state—based on a given metric of interest—from a large ensemble of non-initialized simulations, generating a new ensemble from it, and repeating this process over time to produce decadal predictions. The added value of this approach is that it does not generate any drift. We demonstrate here its application in a case study predicting near-term surface temperatures over the North Atlantic, using observed subpolar gyre sea surface temperature as the basis for member selection.
How to cite: Bonnet, R., Boé, J., and Moine, M.-P.: Development of a new process-based constraint technique to provide decadal climate prediction over Europe, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-18318, https://doi.org/10.5194/egusphere-egu26-18318, 2026.
The winter seasonal variability of the Northern Hemisphere Stratospheric Polar Vortex
(SPV) is characterized by extreme events known as Sudden Stratospheric Warmings
(SSWs). SSWs feature a rapid increase of the temperature of the stratospheric polar
cap and a reversal of the zonal winds, exerting a downward impact on the troposphere
that makes their prediction societally relevant. Yet, despite their importance, current
seasonal prediction systems (SPSs) show only limited skill in forecasting SPV
variability, and SSWs occurrence. This study quantifies to what extent existing
statistical--dynamical approaches, that were designed for the prediction of tropospheric
modes of variability, can improve SPV seasonal predictions. More specifically, we
investigate the efficacy of a teleconnection-based subsampling for the North Atlantic
Oscillation (NAO), which selects a subset of ensemble members based on their NAO
representation. We do this by combining the Copernicus Climate Change Service
(C3S) dynamical forecasts with statistical predictive models for the NAO index. This
approach is motivated by the established link between NAO and stratospheric
variability. Results show that the teleconnection-based subsampling technique
increases the prediction skill of winter SPV variability and of the number of SSW days
in a season, with improvements that are robust across all C3S individual models.
These improvements come together with a better representation of extra-tropical lower-
stratosphere wave activity during December and January (DJ). However, the study
demonstrates that only some of the NAO statistical predictors correlate with SPV
variability, and others show contrasting behavior. This has implications for future
development of hybrid forecast systems targeting both NAO and SPV.
How to cite: Gargiulo, F., Ruggieri, P., Famooss Paolini, L., and Di Sabatino, S.: A hybrid statistical-dynamical approach for seasonal prediction of the boreal winter stratosphere, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-12661, https://doi.org/10.5194/egusphere-egu26-12661, 2026.
Climate change is accelerating the frequency and intensity of heatwaves at an unprecedented rate, posing substantial threats to public health and socio-economic systems. To effectively mitigate future risks associated with extreme heat events, it is crucial to understand the decadal variability of heatwaves and to develop robust medium- to long-term adaptation strategies. However, owing to the complexity of internal climate variability, producing reliable, high-resolution decadal predictions of heatwaves over Taiwan and East Asia remains a major challenge.
This study focuses on the Pacific Meridional Mode (PMM), a key mode of climate variability that influences heatwave activity in East Asia through its modulation of large-scale atmospheric circulation over the North Pacific. By examining the spatial characteristics of heatwaves in Taiwan and their linkage to PMM variability, we develop statistical models that relate PMM to decadal variations in heatwave frequency across East Asia. To further enhance predictive skill and reduce model uncertainty, we also apply a Bayesian ensemble approach, which optimally combines information from multiple models based on their historical performance.
Our results demonstrate that incorporating PMM significantly improves the predictive skill of decadal heatwave forecasts, while the Bayesian ensemble method provides additional gains in forecast accuracy and robustness. These findings highlight the critical role of large-scale climate variability in governing extreme heat events and underscore the value of Bayesian ensemble techniques for advancing decadal climate prediction and supporting proactive climate risk management in Taiwan and East Asia.
How to cite: Tsai, C.-T., Tseng, W.-L., Wang, Y.-C., Shen, Y.-L., and Chu, P.-H.: Linking the Pacific Meridional Mode to Decadal Heatwave Prediction in Taiwan and East Asia, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-15590, https://doi.org/10.5194/egusphere-egu26-15590, 2026.
El Niño–Southern Oscillation (ENSO) variability arises from nonlinear interactions between the tropical ocean and atmosphere, combining deterministic recharge–discharge dynamics with episodic stochastic forcing. While conceptual models such as the Cane–Zebiak recharge oscillator capture the core physics of ENSO, forecast skill remains highly intermittent, with pronounced failures during periods of strong atmospheric variability. This study investigates the dynamical origins and predictability limits of ENSO by integrating a hierarchy of models, ranging from idealised oscillators to observationally forced and hybrid forecast frameworks.
Using ERA-20C equatorial zonal wind anomalies and Niño-3.4 sea surface temperature data, we identify and characterise Westerly Wind Bursts (WWBs) as state-dependent atmospheric perturbations that preferentially occur during warm ENSO phases. When imposed on the Cane–Zebiak oscillator, WWBs act as phase-dependent energy injections, modulating growth and decay without fundamentally altering the underlying oscillatory structure. However, these same perturbations substantially reduce the memory of the system, shortening the energy autocorrelation timescale from approximately 18–24 months to about 6 months during WWB-active periods.
A supervised forecast framework at an 8-month lead time reveals strong regime dependence in predictability. While including atmospheric forcing improves mean forecast skill, performance collapses during WWB months, with large, biased, and phase-dependent errors. Linear and machine-learning residual correction models fail under cross-validation, indicating that WWB-induced errors are not deterministically predictable on an event-by-event basis. Instead, forecast error variance exhibits robust phase dependence, enabling the identification of distinct uncertainty regimes.
Building on this structure, we introduce a hybrid, regime-aware forecasting strategy that applies deterministic prediction only during low-uncertainty regimes and adopts conservative alternatives during high-uncertainty WWB conditions. This approach reduces catastrophic forecast errors and improves reliability without overfitting. Overall, the results demonstrate that ENSO is a conditionally predictable system, where atmospheric forcing not only modulates amplitude but imposes intrinsic limits on deterministic forecast skill. These findings argue for forecast systems that explicitly represent uncertainty and regime transitions rather than relying solely on universal deterministic correction.
How to cite: Pigelet, A.: Understanding Atmospheric Oscillations and Climate Change Using a Hierarchy of Models, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13645, https://doi.org/10.5194/egusphere-egu26-13645, 2026.
Introduction: Observational and modeling studies have examined the
interactions between El Niño-Southern Oscillation (ENSO) and the equatorial
Atlantic variability as incorporated into the classical charge-recharge oscillator
model of ENSO. These studies included the role of the Atlantic in the
predictability of ENSO but assumed stationarity in the relationships, i.e., that
models’ coefficients do not change over time. A recent work by the authors has
challenged the stationarity assumption in the ENSO framework but without
considering the equatorial Atlantic influence on ENSO.
Methods: The present paper addresses the changing relationship between
ENSO and the Atlantic El Niño using an extended version of the recharge
oscillator model. The classical two-variable model of ENSO is extended by
adding a linear coupling on the SST anomalies in the equatorial Atlantic. The
model’s coefficients are computed for different periods. This calculation is
done using two methods to fit the model to the data: (1) the traditional method
(ReOsc), and (2) a novel method (ReOsc+) based on fitting the Fisher’s Z
transform of the auto and cross-correlation functions.
Results: We show that, during the 20th century, the characteristic damping rate
of the SST and thermocline depth anomalies in the Pacific have decreased in
time by a factor of 2 and 3, respectively. Moreover, the damping time of the
ENSO fluctuations has doubled from 10 to 20 months, and the oscillation
period of ENSO has decreased from 60-70 months before the 1960s to 50
months afterward. These two changes have contributed to enhancing ENSO
amplitude. The results also show that correlations between ENSO and the
Atlantic SST strengthened after the 70s and the way in which the impact of the
equatorial Atlantic is added to the internal ENSO variability.
Conclusions: The remote effects of the equatorial Atlantic on ENSO must be
considered in studies of ENSO dynamics and predictability during specific
time-periods. Our results provide further insight into the evolution of the ENSO
dynamics and its coupling to the equatorial Atlantic, as well as an improved tool
to study the coupling of climatic and ecological variables.
How to cite: Cao-García, F. J., Crespo-Miguel, R., Polo, I., Mechoso, C. R., and Rodriguez-Fonseca, B.: ENSO coupling to the equatorialAtlantic: Analysis with an improved recharge oscillator model, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-21543, https://doi.org/10.5194/egusphere-egu26-21543, 2026.
Posters on site: Thu, 7 May, 14:00–15:45 | Hall X5
In recent years, hybrid statistical-dynamical approaches have emerged as a promising avenue to enhance seasonal predictions of the extratropical climate (Slater et al., 2023). Among these, the teleconnection-based subsampling has been shown to significantly improve seasonal predictions of Eurasian climate, including the occurrence of summer extreme temperatures (Famooss Paolini et al., 2024). This technique relies on selecting a subset of ensemble members whose predictions of summer low-frequency atmospheric variability are consistent with its statistical forecasts from springtime predictors.
Here, we assess the potential of the teleconnection-based subsampling to enhance seasonal predictions of two health-related heat-stress indicators in summer: the Universal Thermal Climate Index (UTCI) and the Wet Bulb Globe Temperature (WBGT), which combine information on temperature, humidity, radiation and wind. The methodology is implemented to mimic real-time operational forecast environment, thus differing from standard retrospective forecast (hindcast) applications. We use the ECMWF seasonal prediction system initialised on May 1 and ERA5 reanalysis as surrogate of observations, assessing the prediction skill during 2003—2024. The ECMWF ensemble is subsampled by retaining only those ensemble members that best capture the teleconnection patterns associated with the summer North Atlantic Oscillation (NAO), the leading mode of summer low-frequency atmospheric variability over the North Atlantic sector.
Our results show that the teleconnection-based subsampling in operational forecast environment increases seasonal prediction skill of the summer NAO, moving from near-zero correlation for the full ECMWF ensemble to about 0.45 for the subsampled ECMWF ensemble. In turn, constraining the ensemble to members with a realistic NAO phase enhances the prediction skill of both the UTCI and WBGT, particularly over the Scandinavia and western Europe, where the summer NAO exerts its strongest influence on heat-stress conditions. Results also show that this improvement is mainly linked to a better representation of thermal conditions rather than wind in those regions.
These findings are particularly relevant, as they contribute to the development and implementation of innovative methodologies for predicting climate conditions that pose risks to human health. This is a key priority in the context of climate change, which is projected to substantially increase heat-related mortality unless strong mitigation and adaptation strategies are adopted (Masselot et al., 2025).
Bibliography
Famooss Paolini et al. (2024). Hybrid statistical-dynamical seasonal prediction of summer extreme temperatures in Europe. Quarterly Journal of the Royal Meteorological Society, 151(766). https://doi.org/10.1002/qj.4900
Masselot et al. (2025). Estimating future heat-related and cold-related mortality under climate change, demographic and adaptation scenarios in 854 European cities. Nature Medicine, 1-9. https://doi.org/10.1038/s41591-024-03452-2
Slater et al. (2023). Hybrid forecasting: blending climate predictions with AI models. Hydrology and earth system sciences, 27(9), 1865-1889. https://doi.org/10.5194/hess-27-1865-2023
How to cite: Famooss Paolini, L., Ruggieri, P., Di Napoli, C., Wetterhall, F., Pascale, S., Brattich, E., and Di Sabatino, S.: Enhancing seasonal forecast of health-related heat-stress indicators through teleconnection-based subsampling, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-494, https://doi.org/10.5194/egusphere-egu26-494, 2026.
Northwestern China (NWC) has a monsoon-like, arid and semi-arid climate with considerable decadal variability and long-term trends. Decadal prediction of summer precipitation remains challenging due to the mixed influence of external forcing and internal variability. This study shows that the decadal internal variability of domain-averaged summer precipitation over NWC (NWCP) primarily originates from the extratropical North Atlantic dipole (NAD) sea surface temperature anomalies (SSTA), which excite a Eurasian Rossby wave train by enhancing the transient eddy forcing. The resultant anomalous Mongolian cyclone increases the NWCP through the cyclonic vorticity-generated upward moisture transport. By combining this empirical relationship and dynamical models’ predicted NAD SSTA, we attempted a hybrid dynamic-empirical model to predict the decadal internal variability component. After adding the external forcing component, the model can predict the decadal NWCP 7–10 years in advance. Our result opens a pathway for decadal prediction of precipitation in central Eurasia’s dry regions.
How to cite: Xiang, Y. and Li, J.: Decadal predictability of summer precipitation in Northwestern China originated from the North Atlantic Ocean, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6362, https://doi.org/10.5194/egusphere-egu26-6362, 2026.
Understanding climate and atmospheric teleconnections and the ability of forecasting models in reproducing these with sufficient accuracy represents a key step to better understand model performance and forecasting skill. In this work, we apply the Peter and Clark causal discovery algorithm (PCMCI) to analyse the relationship between the Indian summer monsoon (ISM) and heat extremes in the eastern Mediterranean in ERA5 reanalysis and SEAS5 seasonal forecast data over the 1981-2023 period. The analysis aims to (a) determine the effect of ISM interannual variability on heat extremes, (b) assess the ability of SEAS5 in reproducing the observed causal chain and (c) understand how El Niño affects this teleconnection. Our results show that in ERA5, weak ISM years are connected to an increased probability of heat extremes in Egypt, the Middle East and the Anatolian Peninsula, while in SEAS5 this connection is shown predominantly for Egypt, but it is very weak or absent in the other two regions. Applying PCMCI at sub-seasonal (3-day) time scales shows that SEAS5 can quantitatively well reproduce the causal links connecting the ISM convective activity to the Etesians via the Middle East ridge (ME-ridge). In turn, the Etesians (summer surface northerly winds blowing over the Aegean Sea) affect temperature variability over the region, with weak Etesians leading to higher surface temperature. In contrast, the ability of SEAS5 to reproduce observed causal links diminishes when monthly time scales are analysed. SEAS5 struggles to reproduce the sign of the link from El Niño towards the ISM and from the latter toward eastern Mediterranean geopotential heights. Finally, we assess historical trends and the effect of El Niño on the detected causal links, showing that (a) the effect of the ISM on the ME-ridge has increased since 1981, and that the influence of the ISM on both the ME-ridge and the Etesians is enhanced during La Niña years.
Di Capua et al. “Increased risk of heat extremes in the Eastern Mediterranean during weak Indian summer monsoon years”, in review
Pre-print available at https://www.researchsquare.com/article/rs-6887363/v1
How to cite: Di Capua, G., Zittis, G., Economou, T., Tyrlis, E., and Anagnostopoulou, C.: Using casual discovery to assess the effect of Indian summer monsoon on summer heat extremes in the eastern Mediterranean, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-11486, https://doi.org/10.5194/egusphere-egu26-11486, 2026.
Climate models are plagued by long-standing biases that degrade predictions. While increasing resolution of global climate models to km scales promises to reduce biases, there is little evidence so far of improvements with currently available computing power. Supermodelling is a high-level AI approach that combines existing models with machine learning. This alternative approach has demonstrated reductions in long-standing biases, such as the double ITCZ and tropical SST biases, at a fraction of the computational cost of km scale models. A supermodel is a combination of models that interact during their simulations to mitigate errors before they develop into large-scale biases. Here, I will present recent results from a supermodel based on three Earth System Models (NorESM, CESM, MPIESM). The models were combined using ocean data assimilation and trained on observed SST data. The simulation of tropical climate is markedly improved compared to that of the respective standalone models. We have performed the seasonal predictions using this supermodel and compared them with those from the standalone models. Our results show that while model biases are reduced, seasonal predictions are not necessarily improved. Reduction in biases, however, does lead to improved teleconnections, improving skill over some continental regions.
How to cite: Keenlyside, N., Singh, T., Counillon, F., Schevenhoven, F., Montag, L., and Chiu, P.-G.: Supermodelling as a high-level AI approach, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-12290, https://doi.org/10.5194/egusphere-egu26-12290, 2026.
Seasonal forecasts generated by General Circulation Models (GCMs) provide essential information for early warning and climate services, particularly in regions highly vulnerable to rainfall variability, such as sub-Saharan Africa. To enhance the predictive skill of GCMs, hybrid forecasting systems have been developed that combine physically based dynamical models with data-driven models.
In this study, we apply a statistical-dynamical hybrid method, referred to as teleconnection subsampling, in which AI-based prediction of large-scale teleconnection indices influencing sub-Saharan Africa rainfall -such as the El Niño Southern Oscillation (ENSO), the Indian Ocean Dipole (IOD) and the Atlantic Niño (ATL)- is used as a priori information to select a subsample of GCM ensemble members and generate hybrid rainfall forecast. Previous studies have demonstrated that Convolutional Neural Networks (CNNs) outperform traditional modelling techniques in predicting modes of climate variability; therefore, a CNN-based prediction of a teleconnection index is adopted in this work.
The analysis is based on the ECMWF seasonal forecasting system, provided by the Copernicus Climate Change Service, with ERA5 reanalysis used as the observational reference for skill assessment. The study focuses on three regions -East, West and Southern Africa- characterized by distinct rainfall regimes and lying within the framework of the ALBATROSS (Advancing knowledge for Long-term Benefits and climate Adaptation ThRough hOlistic climate Services and nature-based Solutions) project. The period considered ranges from 1993 to 2016, corresponding to the ECMWF hindcast period. Total precipitation rate and sea surface temperature (SST) fields from ECMWF and ERA5 are used.
Focusing on East Africa, we develop a CNN-based prediction of the IOD index for the October-December season at a lead time of three months. The CNN, trained on Indian Ocean SST anomalies using even years and validated on odd years, is based on the architecture proposed by Tao (2024). The resulting hybrid prediction outperforms both the purely AI-based and purely dynamical predictions and leads to improved rainfall skill, particularly over coastal regions of Kenya and Tanzania.
In addition, experiments assuming perfect knowledge of ENSO and ATL teleconnection indices highlight the potential for further development of CNN-based prediction for June–September rainfall over West Africa, with particularly promising results for the Ghana region. Conversely, limited skill improvements over Southern Africa suggest the need to investigate the role of additional drivers, such as extratropical modes of variability, in future work.
How to cite: Beltrami, S. and Ruggieri, P.: Hybrid seasonal rainfall predictions in sub-Saharan Africa through a teleconnection-based subsampling, informed by AI model, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-12748, https://doi.org/10.5194/egusphere-egu26-12748, 2026.
Climate change adaptation depends on reliable, high-resolution climate data that meet the needs of decision-makers and society. While global climate models (GCMs) provide essential information, they contain systematic biases that must be corrected before use at local scales. This study evaluates statistical bias-adjustment methods applied to temperature and precipitation from the NASA Earth Exchange Global Daily Downscaled Projections (NEX-GDDP-CMIP6) and assesses their suitability for deriving hydro-agro-climatic indicators for Iceland.
The analysis uses data from 12–14 climate models covering a historical period (1950–2014) and a future period (2015–2100) under three Shared Socioeconomic Pathways (SSP2-4.5, SSP3-7.0, SSP5-8.5). The high-resolution (2.5 km) CARRA reanalysis dataset serves as the reference for bias adjustment. For near-surface air temperature, six methods were evaluated: linear scaling; empirical quantile mapping (EQM) with constant, monthly-varying, and 31-day moving-window adjustment factors; and two trend-preserving approaches, detrended quantile mapping (DQM) and quantile delta mapping (QDM). For precipitation, EQM with a 31-day moving window and frequency adaptation, as well as QDM, were assessed.
Results show that trend-preserving methods perform best for bias-adjusting temperature, as they maintain long-term climate signals while reducing systematic errors. For daily precipitation, EQM outperforms QDM, particularly in correcting high-intensity events, reducing biases at the upper 95th percentile more effectively. QDM was less successful in reducing precipitation biases and did not substantially improve trends.
Bias-adjusted projections indicate that Iceland will experience a temperature increase of approximately 2.4–3.1 °C [0.6–4.8 °C] by the end of the century (2071–2100) relative to 1981–2010, depending on the emissions scenario, with the strongest warming in northern regions. Precipitation is projected to increase by more than 2% per degree of warming, with larger increases in autumn than in winter. Annual maximum 24-hour precipitation is expected to rise by 6–7% by mid-century and by 6–14% by late century, corresponding to increases of 3–7 mm per day, with the largest changes under the high-emissions scenario (SSP5-8.5). Extreme precipitation events will become more frequent, with 100-year events potentially occurring three to four times more often under high emissions.
Regionally, southern and southeastern Iceland are projected to become drier, while northern Iceland becomes wetter. This work, conducted as part of the Climate Atlas of Iceland, provides high-resolution, bias-adjusted climate data for national-level impact and adaptation studies and highlights the strengths and limitations of commonly used bias-adjustment methods.
How to cite: Ólafsdóttir, A. H., Zaqout, T., and Björnsson, H.: Bias adjustment of the NEX-GDDP-CMIP6 climate data and predicted change in the future climate of Iceland , EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-13995, https://doi.org/10.5194/egusphere-egu26-13995, 2026.
High-resolution meteorological fields are essential for assessing localized impacts of climate change. In recent years, deep learning (DL)-based downscaling techniques have emerged as a computationally efficient and sustainable alternative to dynamical downscaling, demonstrating strong skill in reconstructing fine-scale features and complex flow characteristics.
Despite these advances, the spatio-temporal physical consistency of DL-downscaled fields remains a critical concern. Because most machine learning (ML) approaches are not explicitly constrained by physics-based equations, their ability to respect fundamental atmospheric balances is still debated. Similar concerns apply to global data-driven forecasting models, which have revolutionized medium-range weather prediction in recent years but often operate as "black boxes" [1].
While the physical integrity of global ML forecasting models has started to receive attention, high-resolution downscaling applications remain largely unexplored from this perspective. In this study, we address this gap by assessing the physical consistency of a Latent Diffusion Model (LDM) - based on the architecture developed by Tomasi et al. 2025 [2] - trained to downscale ERA5 [3] over the Italian peninsula using CERRA [4] as the target dataset.
Moving beyond standard statistical error metrics, we evaluate the model using a suite of physical diagnostic constraints, with particular emphasis on mass conservation and thermodynamic relationships between temperature and moisture. Our results provide a benchmark for the physical reliability of DL-downscaling techniques in regional climate applications, thereby enhancing their credibility and facilitating their broader integration within the atmospheric sciences.
[1] Hakim, G. J., and S. Masanam, 2024: Dynamical Tests of a Deep Learning Weather Prediction Model. Artif. Intell. Earth Syst., 3, e230090, https://doi.org/10.1175/AIES-D-23-0090.1.
[2] Tomasi, E., Franch, G., and Cristoforetti, M.: Can AI be enabled to perform dynamical downscaling? A latent diffusion model to mimic kilometer-scale COSMO5.0_CLM9 simulations, Geosci. Model Dev., 18, 2051–2078, https://doi.org/10.5194/gmd-18-2051-2025, 2025.
[3] Hersbach H, Bell B, Berrisford P, et al. The ERA5 global reanalysis. Q J R Meteorol Soc. 2020; 146: 1999–2049. https://doi.org/10.1002/qj.3803
[4] Ridal, M., Bazile, E., Le Moigne, P., Randriamampianina, R., Schimanke, S., Andrae, U., et al. (2024) CERRA, the Copernicus European Regional Reanalysis system. Quarterly Journal of the Royal Meteorological Society, 150(763), 3385–3411. https://doi.org/10.1002/qj.4764
How to cite: Iacomino, C., Tomasi, E., Franch, G., Cristoforetti, M., and Bordoni, S.: Physical consistency of high-resolution meteorological fields from deep learning-based downscaling, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-16795, https://doi.org/10.5194/egusphere-egu26-16795, 2026.
The Indian Summer Monsoon (ISM) is an outcome of the Intertropical Convergence
Zone (ITCZ) which drives the subcontinent’s economy. There is a growing interest in
investigating the causal mechanisms underlying the monsoon onset. The positive SSH
(Sea Surface Height) anomalies over the West Tropical Indian Ocean (WTIO, 50 to 70°E
& 10°N to 10°S) and the core LLJ zone (low-level south-westerly jets) in the Arabian Sea
are significant contributors of an early monsoon onset. The thermosteric height difference
in the SSH values between the WTIO and the South-east Tropical Indian Ocean (SETIO,
90 to 110°E & 10°S to 0°N) provides an insight to statistically derive causal linkages
to understand how the Precipitable Water Vapor (PWV) is being affected by convective
patterns, which further explains the thresholds and relationships with respect to the
SSH values to accurately ascertain the start of ISM. The intrusion of easterlies into
WTIO from SETIO further tantamount to incremental evaporation owing to carriage of
comparatively warmer sea-water during the late pre-monsoon period (March-April-May),
and thus, have to be causally analysed to introspect its long-term effect in contributing
to the start of ISM over the south-west coast of Kerala (India). In this study, we perform
the causal analysis between thermosteric heights and ISM precipitation to understand the
onset of ISM. The PCMCI+ algorithm gives the causal strengths and lead-lag connections
between the thermosteric component of the SSH values and the ISM onsets to statistically
determine the parent cause behind early/delayed monsoons, emphasizing over the above-mentioned
region of interests. This gives a perception over the inconsistent active and
break phases in the ISM onset patterns over the mainland owing to the variabilities in the
hydrometeorological patterns that adversely affect the Cumulonimbus cloud formation,
and are mainly responsible for the monsoon rainfall over the major climate classes of
India.
How to cite: Mahmood, J. and Devaraju, B.: Can Thermosteric Heights be used as causal indicators of Indian Summer Monsoon Onset?, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-19046, https://doi.org/10.5194/egusphere-egu26-19046, 2026.
